[Beowulf] new mpich/pvfs2

[Michael Gauckler]

Dear list, dear mpich/pvfs2 developers,

thank you very much for the Release of mpich2 version 1.0.2 together
with PVFS2 1.1.0. We noticed that the performance increased and is now
independent of the cb_buffer_size. We are experiencing datarates of up
to 220 MB/s (which was previously only achievable for certain sizes of
cb_buffer_size) measured by coll_perf. Also memory utilisation when
flattening the MPI_Type_create_subarray came down (from in our case
~0.5 GB).

To write a 1.6 GB file from 2 processes in a subarray-sliced manner to
disk and read it again, 191 sec. are needed (~17 MB/s, see [1]). Doing
the same test with 4 processors (3.2 GB of data) unfortunately is not
possible since the program crashes. Before the upgrade of mpich/pvfs2
the programm run with 4 processors, so the problem is unlikely in the
program (see [2]) itself.  The debugger stops the program in a PVFS
call (see [3])

We are happy for any hint how to avoid this or any
timings/confirmation from people who are willing to test our little
program on their system.

Thank you for your time!

Yours,
 Michael

__

We are running Debian on Intel Xeon:
Linux 2.6.8-2-686-smp #1 SMP Thu May 19 17:27:55 JST 2005 i686
GNU/Linux


[1]
time mpdrun -1 -np 2 ./mpicube2
real    3m11.262s
~17 MB/s

[2] --> Attachement "mpicube2_mpich_v102_pvfs2_v110.cpp"
(See attached file: mpicube2_mpich_v102_pvfs2_v110.cpp)

[3] --> http://www.parallelsystems.ch/mpicube2_debug.png

-------------- next part --------------
A non-text attachment was scrubbed...
Name: mpicube2_mpich_v102_pvfs2_v110.cpp
Type: text/x-c++src
Size: 9890 bytes
Desc: not available
URL: <http://www.beowulf.org/pipermail/beowulf/attachments/20050612/444ee61e/attachment.cpp>